Linear optimal prediction and innovations representations of hidden Markov models

The topic of this paper is linear optimal prediction of hidden Markov models (HMMs) and innovations representations of HMMs. Our interest in these topics primarily arise from subspace estimation methods, which are intrinsically linked to such representations. For HMMs, derivation of innovations representations is complicated by non-minimality of the corresponding state space representations, and requires the solution of algebraic Riccati equations under non-minimality assumptions.